Add predictive modeling of the interests and creators a user will engage with next. Train with multi-objective goals (engagement, satisfaction). Redesign the interest vocabulary for much higher precision, and evaluate learned alternatives to today's interest vocab. Build deep, sequence-based token models that learn from long user activity histories across Google products. Build offline evaluation that predicts online results (profile-quality autoraters, retrieval and ranking metrics), and run well-sliced live experiments. Design a shared token description, feature set, and training objectives across token types.
Job description provided by Google.
View the original posting
